Profile Attributes As Similarities

Howard R. Moskowitz, Pioneering Research Laboratory, U.S. Army Natick Laboratories
[ to cite ]:
Howard R. Moskowitz (1974) ,"Profile Attributes As Similarities", in NA - Advances in Consumer Research Volume 01, eds. Scott Ward and Peter Wright, Ann Abor, MI : Association for Consumer Research, Pages: 182-191.

Advances in Consumer Research Volume 1, 1974    Pages 182-191

PROFILE ATTRIBUTES AS SIMILARITIES

Howard R. Moskowitz, Pioneering Research Laboratory, U.S. Army Natick Laboratories

INTRODUCTION

In food science and in the chemical senses (taste and smell) increasing use in being made of multidimensional scaling to represent nuances of flavor, and the relations between chemicals and descriptors. Potentially thousands of different flavor nuances are possible (Harper, 1973), and a principal task for a psychology of flavor is to develop a simple system of descriptors that permits the investigator to see, at a single glance, relations among the products (i.e., foods or simple stimuli) that he tests. In the science of taste and smell the problems are different. Experimenters often are interested in the qualitative characteristics of stimuli in order to understand the underlying mechanism of taste and smell physiology, or the mechanism of taste and smell perception.

Multidimensional scaling has been used recently to represent qualitative similarity among odors (Berglund et al., 1972; Woskow, 1968; Yoshida, 1964; Yoshida, 1972), and tastes (Moskowitz, 1972; Schiffmann & Erickson, 1971; Yoshida, 1963; Yoshida & Saito, 1969). The techniques are subsumed under the rubric of proximities analysis. Observers are asked to estimate the degree of qualitative similarity (or dissimilarity) between pairs of stimuli, and then these ratings are tested by one or another method to yield psychological distances between stimuli. Subsequent analysis of the distances by one of a variety of algorithms produces a spatial configuration of points in which inter-point distances correspond to psychological dissimilarity of stimulus pairs.

A parallel trend to capture quality is known as profiling, and has received wide interest by those interested in the practical aspects of food science. Profiling requires that the experimenter select a series of qualitative descriptors that he feels are relevant for a particular flavor or aroma. Profiles differ in their constitution from study to study, and the specific descriptors may be drawn from hundreds of different words that are available (Harper, 1973). Because of time limits, and redundancy in some descriptions, usually 50 or fewer are used (Harper et al., 1968, von Sydow et al., 1970). The respondent is instructed to sample the stimulus (either by sipping or by sniffing) and then to rate the stimulus with the scale (usually a category scale, containing between 5 and 10 categories of increasing magnitude). Each stimulus is rated on each descriptor, and the mean ratings across a group of observers produce a characteristic profile. The same food undergoing different processes can be compared to a reference profile in order to determine changes in sensory quality. In contrast to the n(n-1)/2 comparisons for dissimilarity required by proximity analyses, profiling n stimuli on m descriptors calls for only mn comparisons.

Profiling and proximities analyses provide complementary information. Profiling yields an image of the sensory constitution of a food when the observer assumes an analytic role. The dimensions that the experimenter selects to be used for the profile are critical, and the profile may not fully characterize the food being rated. Proximities analysis, in contrast, requires that the observer generate a series of operating rules by which he makes dissimilarities judgments, and does not impose any requirement on the experimenter to pre-select dimensions that might be relevant. Often dimensions arising from proximities data are not immediately or directly interpretable, and the entire array of points in the geometrical space may change position relative to each other when yet another point or set of points is evaluated. The solution from proximities analysis may be applicable only to the array of stimuli that are evaluated together.

Recently, profiling and proximities analyses have been combined to yield a 'joint space' containing both chemical stimuli and descriptors (Alabran & Moskowitz, 1973; Moskowitz & Gerbers, 1973). Conceptually, the procedure is straight forward. A profile, in which the entries are magnitudes or intensities is considered to be a matrix of proximities between words (descriptors, ideal points) and stimuli (odors or flavor stimuli). The higher the rating the more 'similar' are the word and the stimulus. This approach is a variety of unfolding analysis, in which words become ideal points, and stimuli become points compared to the ideals. On occasion, when profile ratings are obtained across a series of increasingly intense stimuli (viz. the sensory constitution of a concentration series of amino acids or other complex-tasting stimuli) the profile ratings for each stimulus should be normalized, so that the total intensity becomes 1.0 (or some other constant). If a stimulus is rated only on one dimension, then that dimension receives a 1.0, and all others are 0.0. If the stimulus is equally strong along two dimensions, then each is rated 0.5, etc.

The present paper presents a re-analysis of three sets of profiling data, considered in terms of the unfolding paradigm. They are:

1) Evaluation of the aroma of canned and processed beef

2) Evaluation of the aroma constitution of bilberries

3) Evaluation of the odor constitution of 15 odorants rated on four separate days of experimentation.

These three data sets were analyzed by the multidimensional scaling program MDSCAL 5M (Kruskal & Carmone, 1969), with the stipulation that high entries for the profile matrix corresponded to similarities. Solutions were obtained in dimensions 3, 2 and 1, but no rotation was made to any specified target configuration.

RESULTS

Figure 1 shows the two dimensional flavor space obtained from the reanalysis of data from profiles of meat undergoing different treatments (stress for the unfolded profile = 0.13). The collection of meats with various treatments from an elongated cluster, lozenge shaped, whose closest reference descriptors are 'cooked meat' and 'preference'. Since the meat stimuli lie so close to each other, it is probable that substantially more variation occurred for ratings of the same meat on different descriptors than between meats. Otherwise, the meats would be qualitatively more dissimilar, and would be expected to distribute in the space just as descriptors do.

It is possible to use the two dimensional space to trace out contours of equal dissimilarity (or equal similarity) ratings. The distance metric is the Euclidean distance, so that a fixed distance can be swung around 360¦, with the reference point being 'preference' (i.e., determination of 'iso-preference' difference contours) or 'sour' (i.e., iso-sourness difference contours) etc. This exercise is done most profitably by starting somewhere in the-space that is densely packed with stimuli and with descriptors, and it can illustrate relations among descriptors which could not have been previously determined.

As distances of descriptors from meats increase, the descriptor becomes less important for characterizing the food. For instance of intermediate importance are the descriptors 'cooked vegetables', 'cooked cabbage' and 'sulfurous', all of which have sulfurous notes. Finally, the least relevant for preferred products are 'sweet', 'metallic', 'sour', etc. Were some unacceptable meat samples to be included in the series, which are qualitatively different from the collection evaluated by Persson et al., (1973) the unimportant descriptors might take on more importance as indicators of less preferred stimuli.

Figure 2 shows a reanalysis of the profile data for bilberry juice, collected by von Sydow, (1970). In contrast to the two dimensional configuration shown for meats, bilberry aroma evaluated across a variety of different samples comprises principally a unilateral continuum, as determined from unfolding. The degenerate configuration yields two clusters of descriptors, and one cluster of stimuli. One of the important outcomes of this analysis is the recognition that for an appropriate joint space to be obtained by unfolding, a variety of qualitatively different foods must be evaluated by the profile procedure.

Figure 3 shows a two dimensional representation of odor profiles obtained for 15 different odorants rated of four separate days on a series of 16 attributes. Each attribute was assumed to remain invariant in its meaning across the four days, but each odorant rated on days 1 - 4 was treated as a separate stimulus. This yielded a matrix of 60 (15 odorants x 4 days) x 16 (descriptors) (Moskowitz & Gerbers, 1973). The descriptors and the odorants are both embedded in the space, and for each odorant a number represents the ordinal day when it was evaluated.

The advantage of unfolding the profile is that the qualitative nuances of odors can be compared to each other from one day to another, by means of 'triangulating' odors against an invariant set of descriptors (or ideal points). Odorants that do not move about in the space, and whose position lies fixed from one day (or session) to the next are those whose profiles remains substantially unaffected by experience. Those odorants that do move substantially are either a) odorants whose quality is perceived to change with experience (for example, if the observer notices new emergent notes as he becomes familiar with the smell, or if the odor is biologically determined by the interaction of chemicals and living organisms, as in cheese) or b) odorants that lie on an edge of a three (or higher) dimensional solid that is projected for convenience onto two dimensions. The second possibility, b, means that the shift is artifactual, and results from the difficulties encountered in projecting a figure of high dimensionality to one of low dimensionality while simultaneously striving to maintain the topological relations.

In Figure 3 the three dimensional configuration (not shown) for the unfolded data is governed by a stress of 0.45, whereas the two dimensional configuration is governed by a stress of 0.51, and the one dimensional configuration by a stress of 0.75. Hence, possibility b is feasible. Previous results, however, with the same odors (Moskowitz & Gerbers, 1973), with direct judgment of dissimilarity between all pairs of odorants and stimulus words (total = 900 pairs) suggest that some odors, such as xylol, characteristically shift in position across the four days of experimentation. Others, such as limonene (which has the strong character note of line) shift less dramatically.

It is quite possible that for some odorants repeated experience, possibly coupled with the ability of the observer to develop labels for odorants and to see new characteristic quality, will yield an altered impression of olfactory quality. It is the unique ability of profiling and unfolding analysis to uncover such perceptual shifts of odor quality that make repeated profiling with the same odorants across days very attractive as an exploratory tool for olfactory science and food technology.

DISCUSSION

The use of the joint space of descriptors and stimuli (whether pure chemicals or foods) allows the word descriptors to be evaluated via their correlated use in products, as well as products evaluated via their correlated use in word descriptors. The advantage of this procedure, as noted before, is the development of the joint space. Distances between words and chemicals (foods) thus reflect distance between a stimulus and an ideal point.

Quite often the odor descriptors lie at the edges of the space. This implies that odor stimuli tend to be more similar to each other than descriptors are to other descriptors, and that descriptors are more 'saturated' with odor quality (in analogy to color science) than are odorants. It would be instructive in future studies to develop unfolded profiles for a variety of food descriptors as well as those actual foods. This proposed experiment would indicate whether the descriptor of a food possesses 'more' of the quality than a physical specimen of the food itself. In a previous study (Moskowitz & Gerbers, 1973) the same type of saturation phenomenon occurred, with the exception of two descriptors 'heavy' and 'pungent', which were located towards the center of the space.

Several advantages in flavor and odor work accrue from using multidimensional scaling, and especially from the unfolding procedures described here. First, odorants can be directly compared to descriptors on an intensity rating scale. In flavor work with panels the rating procedure is well accepted, and with just a little bit of training the profiling is easily done. Since the sessions can be shortened or lengthened for convenience, unfolding of profiles provides almost comparable information to proximities analyses, but as costs amendable to an applied research organization.

Second, since odorants are directly compared to descriptors, and not to each other, one need not worry about the shifting of attention of the panelists from one set of salient aspects to another as a function of what stimuli are being compared to each other.

Third, traditional multidimensional scaling requires that the observer determine his own rules for proximities judgments, and seeks to uncover those rules through multidimensional placement of points in a space. Unfolding analysis, in contrast, specifies as far as possible the rules by which the observer is to make his judgments by instructing him on the qualities to rate. This 'rigidity' can become important when unfolding is used in quality control of products assessed by the profiling procedure, or when a variety of different products are to be evaluated by different groups of observers, and subsequently compared together in one analysis.

Fourth, in the evaluation of flavor stimuli and perishable foods, storage time and other variables affect quality. Proximities analysis is impossible for the same product separated by storage times of weeks, months or even years. Profiling analysis against invariant verbal concepts or against reference chemicals becomes more feasible. For example, if a wine is sampled in one year, and left to age two or three years, the vintner can determine the course of flavor development, and the optimum storage time. Withdrawal tests of foods over time (undergoing long-term storage stability tests to evaluate shelf-life) can employ profiling and unfolding analyses to determine qualitative shifts that might eventually render the product unacceptable in flavor.

Fifth, profiles can be analyzed across observers to yield an observer x stimulus space. The degree to which the observers differ from each other in their respective spaces indicates the perceptual differences among individuals. Obtained at any one time, comparative 'perceptual maps' of different observers provide only information about the variation in the population of individuals. In training procedures, wherein the aim is to teach the panelists to respond in a standardized way (i.e., to 'calibrate' the observers), the perceptual maps uncovered by unfolding the profiles of the same foods across training days illustrates the convergence of the panel towards the desired congruence in judgments.

Finally, profiling procedures, like any other multidimensional scaling method, is useful for the evaluation of psychophysical processes. Quite often the finer nuances of a taste or odor mixture are evaluated simply on the basis of hedonic or intensitive shifts. Rarely are qualitative shifts in odor or flavor quantified and probed in depth in psychological studies in the chemical senses. With profiling procedures, a large number of mixtures of stimuli can be evaluated simultaneously against a group of standards. The components of the mixture can be evaluated against the same standards as well. Each point in the space would then represent a stimulus (either a simple, unmixed stimulus or a mixture of two or more stimuli). If there exists a predictive and 'additive' system for odor mixtures, based upon the sensory qualities of their components, then mixtures would be expected to lie on the line (or convex region) connecting their components. By embedding both types of stimuli (simple and mixtures) in the same space, departures from simple mixture phenomena can not only be assessed, but the direction of shift can be evaluated both in terms of stimuli that make up the mixture, and in terms of the semantic descriptions given to the mixture. For instance, if in a mixture of a fragrant odorant (e.g., isobutyl isobutyrate) and a repulsive one (e.g., methyl disulfide), the odorant is rated more 'goaty', more 'burnt', and more 'sharp' than either component, this fact would show up immediately on the geometrical representation. Although this final suggestion is by way of a data analytic approach, it may well pave the way for new approaches into the psychology of odor (and taste) mixtures.

FIGURE 1

UNFOLDED BEEF PROFILES

FIGURE 2

UNFOLDED PROFILES OF BILBERRY JUICE

FIGURE 3

UNFOLDED ODORANT PROFILES ACROSS DAYS

FIGURE LEGENDS

Figure 1: Unfolded, two dimensional maps for the flavor of beef, obtained from a reanalysis of the data of Persson, von Sydow & Akesson (1973). Profile descriptors are capitalized. Letters and numbers refer to beef treatments as follows: N,M = canned beef heated to 121¦C for 30 minutes, and evaluated for aroma and taste by nose and by mouth, respectively; 1,2 = canned beef and canned beef + fat + carbohydrate, respectively, each evaluated for aroma intensity without a standard reference sample being presented to the subject; 3,4 = canned beef and canned beef + fat + carbohydrate, respectively, each evaluated for aroma intensity with a standard reference sample for comparison; 5,6,7,8 = canned beef heated for 30 minutes at 131¦C, and evaluated for aroma. The formulations were beef + fat (5), beef alone (6), beef + carbohydrate (7), and beef + fat + carbohydrate (85.

Figure 2: Unfolded, 1 dimensional profile for the aroma of bilberry juice, obtained from a reanalysis of the data of von Sydow et al., (1970).

Figure 3: Unfolded, two dimensional profile of 15 odors evaluated on four successive days along a profile of 16 descriptive terms. The chemicals are abbreviated, and the number at the end indicates the day on which the ratings were taken The odorants were treated as a set of 60 different chemicals. The abbreviations are ACE = acetophenone, BEN = benzaldehyde, BUTAC = butyl acetate, CAM = camphor, CAP = caproic acid, EUG = eugenol, GUA = guaiacol, ISOB = isobutyl isobutyrate, ISOPR = isopropanol, LIM = limonene, MDS = methyl disulfide, MSA = methyl salicylate, PENT = pentanol, SAF = safrole, XYL = xylol.

REFERENCES

Alabran, D., E Moskowitz, H. R. Unpublished data, 1973.

Berglund, B., Berglund, U., Engen, T., & Ekman, G. Multidimensional scaling of twenty-one odors. Report 345, Psychological Laboratories, University of Stockholm, Sweden, 1972.

Harper, R. The language of food and drink. Proc. Inst. Food Sci. & Technol., In press.

Harper, R., Land D. G., Griffiths, N. M., & Bate-Smith, E. C. Odour qualities: A glossary of usage. Brit. J. Psychol.,1968, 59, 231-252.

Kruskal, J., & Carmone, F. How to use MDSCAL 5M and other useful information. Unpublished manuscript, Bell Laboratories, Murray Hill, N. J., 1969.

Moskowitz, H. R. Perceptual attributes of the taste of sugars. J. Food Science, 1972.

Moskowitz, H. R., & Gerbers, C. Dimensional salience of odors. Annals of the New York Academy of Science, In press, 1973.

Persson, T., von Sydow, E., & Akesson, C. Aroma of canned beef: Sensory properties. J. Food Sci., 1973, 38, 386-392.

Schiffmann, S. S., & Erickson, R. P. A psychophysical model for gustatory quality. Physiol. & Behav., 1971, 7, 617-633.

von Sydow, E., Andersson, J., Anjou, K., Karlsson, Land, D., & Griffiths, N. The aroma of bilberries (Vaccinium myrtillus L.) II. Evaluation of the press juice by sensory methods and by gas chromatography and mass spectrometry. Lebensm. Wiss. und Technol., 1970, 3, 11-17.

Woskow, M. H. Multidimensional scaling of odors. In: Theories of Odor and Odor Measurement (ed. N. Tanyolac), Circa Publications, Inc., N. Y. 1968, 147-186.

Yoshida, M. Similarity among different kinds of tastes near the threshold concentration. Jap. J. Psychology, 1963, 34, 25-35.

Yoshida, M. Studies in psychometric classification of odors. Japanese Psychological Research, 1964, 6, 11-124, 145-154.

Yoshida, M. Psychometric classification of odors (6). Japanese Psychological Research, 1972, 14, 70-86.

Yoshida, M., & Saito, S. Multidimensional scaling of taste of amino acids. Japanese Psychological Research, 1969, 11, 149-166.

----------------------------------------